gwern

> In a world chasing ever-bigger models, small ones are quietly doing more with less—and that's exactly what makes them powerful.

100%. It has enough technical details that maybe a human did something. But who knows.

On a stylistic note, I was surprised to see David Friedman letting ChatGPT write his posts when he is a perfectly competent writer himself, and then I realized this is Dave Friedman. Turns out to be rather different!

I did skim it, but it didn't help me much, because I didn't see any comparison to the supposed other cohorts, and even if they had done those, an online convenience sample is almost certainly going to differ a lot in many ways from whatever other white cohort you used from a different study, whereas you could've simply recruited white participants from the same service and gotten an obviously much more valid comparison group.

A better question would be, why would you do an online survey of only black men, if your goal is understand "the unique challenges of brain health and cognitive decline faced by Black men"? Surely you would want to leave it open to at least white and hispanic men too, to at least try and see if those challenges are in fact 'unique' and have some sort of baseline or comparison? And it was an online survey of self-reported measures recruited by an existing survey service (CloudResearch), so they were not limited to an existing sample of only black men - they went out of their way to restrict the sample.

Have you noticed the most recent one, gemini-2.5-pro-0506, suddenly being a lot more sycophantic than gemini-2.5-pro-0325? I was using it to beta-read and improve a story (https://news.ycombinator.com/item?id=43998269), and when Google flipped the switch, suddenly 2.5 was burbling to me about how wonderful and rich it was and a smashing literary success and I could've sworn I was suddenly reading 4o output. Disconcerting. And the AI Studio web interface doesn't seem to let you switch back to -0325, either... (Admittedly, it's free.)

OP's workflow might be much more efficient with use of https://github.com/gildas-lormeau/SingleFile/

It can handle most of what they describe for things like private/paywalled pages or media enclosures or completely self-contained archives that live locally or easy to use or editing before saving or ensuring lazy-loaded images are there, you can view it immediately to check for breakage, it automatically works with adblock and NoScript and when you delete stuff in the DOM using the picker so they can clean each page very efficiently (create a bunch of rules in your adblock by picking elements like in ublock, so you never have to do those again, then quickly mouse any remainder), and it stores the final DOM so you can interact with stuff to make sure it is visible or archived.

So what I do ( https://gwern.net/archiving#preemptive-local-archiving ) is I have a script which calls SingleFile-CLI in a headless Chrome browser to automatically archive everything, and then opens up the original URL + snapshot in my normal Firefox, and look at the snapshot then original. If the snapshot looks good, I simply close the 2 tabs after a few seconds and I'm done; if the snapshot looks bad, then I look at the original and make edits: use Ublock Origin to define any necessary rules (assuming the page isn't cleaned up by all the rules I previously defined), make any minor tweaks to the DOM, and then SingleFile-browser-extension it manually.

If you use enough adblock rules, then you get a similar effect to the 'templates' described, since it looks like OP is mostly just trying to remove as much as possible. But since you're archiving the final DOM, you can do anything you like. Something I've done a few times is opening up multiple pages and copy-pasting the key DOM node from each of them into the first one, to create a single consolidated master page, in a way which is a lot easier & more reliable than messing around with the serialized HTML in Emacs.

You can also post-process them. (Because we use these local archives for 'previews' on Gwern.net, and a fully static self-contained HTML page can easily be 100MB+ with all its fonts and images and stuff, we take the SingleFile snapshots and for the large ones, we 'split' them back up, so loading the .html file doesn't necessarily load everything else: https://github.com/gwern/gwern.net/blob/master/build/deconst... And then you can save a lot of space by running standard optimization tools on the split-out files, eg OptiPNG on the revealed PNGs will save gigabytes of space because so many people fail to do the standard image optimizations.)

Compared to "it typically takes me a few minutes to save a page", I handle the majority of pages in a few seconds, and even the nastiest page where I have to delete a lot is usually like a minute. And since I do like 10 URLs a day, this is quite manageable at scale. (I'm up to >15k snapshots, although an unknown fraction are from an initial bulk archiving so may not be of high quality.)

Those studies never 'control for those'. What they do is a crude statistical approximation (sometimes extremely crude - eg most studies will 'control for education' by counting up 'years', which equates 4 years at Caltech with 4 years in community college), and hope that not too much leaks through as "residual confounding" (https://journals.plos.org/plosone/article?id=10.1371/journal...). Unfortunately, because everything is correlated, there's residual confounding everywhere. This is why every time someone does a Scandinavian population registry study and compares eg siblings within the same family, often the correlation just disappears then and there.

It's 100% unsurprising that this is true of fitness too. This is what always happens. You look at something like a corporate health fitness plan and you find some correlate even after you 'control for' SES, prior health record etc etc; wonderful! Then you do a randomized experiment and it turns out that the residual confounding was still larger than any causal effect which might be there: https://gwern.net/doc/statistics/causality/2022-wallace.pdf Ah well. Maybe next time you'll manage to 'control for' the confounders...

> Work in progress (or 404)

I think it would be a lot politer if the target page was clearer. As it is, I could be 'voting' on a draft or it could just be plain broken and the intended page somewhere else.

Personally, I think it would make more sense to just put up what you have explicitly noted as drafts etc, and count the 'votes' that way.

> It felt like the first 3 (or 2.5) paragraphs, which were arguing that Bjork needed an official website, were a bit of a tangent from the main argument of the article, which was that we need more professional critics, but social media has essentially defunded and dethroned them.

Not so much of a tangent as just the relevant argument not being made clearly. The Bjork example demonstrates the value of a central, canonical source for information in overcoming the costs of friction from direct messaging, which creates a chaotic cacophony of tiny bite-sized messages which are difficult and exhausting to piece together into a final meaningful message, and result in the interested Bjork fans living in their own little information-universes: in one universe, it's a film+documentary, in another, it's a film. So they can't even manage to agree on the most basic facts. (Which has downstream effects: a Bjork fan may not know they have access to the documentary or that they can assume most of the film-watchers saw the documentary and they can invoke it without confusion or spoilers.) The 'advantage' of social media and disintermediation proved to be illusory as they came with too much overhead and destruction of any canon or commons.

It would be more convincing if they did an exhaustive enumeration and verified that for every possible 3x3 Life the learned NN was correct. How do I know looking at a speckled screenshot that it is exactly correct and there's not a little floating point error somewhere or something like that which results in 1 edge-case being slightly off? If the only testing is '100 Life games for 100 steps', that isn't water-tight. (While if you do exhaustive enumeration, well, it has to be correct, because the NN is deterministic and fixed and there's no way for it to go wrong then.)

(I wondered if I overdid it, but I was also thinking that it's necessary to overdo it at least slightly because that's the fake-out ending, and you don't want it to slide by the reader unnoticed. "And everyone thought the meeting of the man and woman had been by chance"...)

There are a number of variations. You might actually be thinking of https://www.overcomingbias.com/p/my-favorite-liahtml or possibly https://gwern.net/doc/statistics/prediction/2022-gelman.pdf (if neither of those are it, it might be one of the others I collated in https://gwern.net/fake-journal-club#external-links ).

Relevant recent discussion: https://news.ycombinator.com/item?id=43736172

Balatro was started in vacation time and underwent a ton of tweaking: https://localthunk.com/blog/balatro-timeline-3aarh So if it had to be written in C, probably neither of those would have happened.

> has been externalized to customers in search of ill-gotten profits.

'Externality' does not mean 'thing I dislike'. If it is the customers running the software or waiting the extra couple of seconds, that's not an externality. By definition. (WP: "In economics, an externality is an indirect cost (external cost) or benefit (external benefit) to an uninvolved third party that arises as an effect of another party's (or parties') activity.") That is just the customers picking their preferred point on the tradeoff curves.

A subtext here may be his current AI work. In OP, Carmack is arguing, essentially, that 'software is slow because good smart devs are expensive and we don't want to pay for them to optimize code and systems end-to-end as there are bigger fish to fry'. So, an implication here is that if good smart devs suddenly got very cheap, then you might see a lot of software suddenly get very fast, as everyone might choose to purchase them and spend them on optimization. And why might good smart devs become suddenly available for cheap?

Note that he did not use the expression you are criticizing him for misusing.

The rule is the rule, and exceptions are the exception. Exceptions do not make the rule, by definition, so if your only defense of Brutalism is to say 'look at this one exception out of the tens of thousands that got built, which doesn't suck!', then you have conceded the point about Brutalism sucking.

It is pretty much the same scaling, though: https://arxiv.org/pdf/2412.09871#page=10 It just lets you avoid some of the pathologies of BPEs.

One wonders how much of it is true and how much is fundamental attribution bias.

> Are you implying that by smoking I caused irreversible changes to my brain

Yes. Once an addict, always an addict. ('Irreversible changes to a brain' are quite common. You remember having smoked, for example...) More importantly, this was something emphasized before, and so it is irrelevant to bring it up as a supposed counterexample.

The key is, they are differently unstable. Two unstable careers can combine to a stabler career. While you work on the novel, you collect the check for the video game, say.

The challenge always is ensuring that when you serve two masters, you don't wind up serving only one (as happens to many novelists or writers who went to Hollywood, and perhaps now with video game projects...).

> I did and while I like Gwern's writing, I think in this case it's plain wrong. I say this as a smoker who switched to e-cigarettes for a few years and then quit cold turkey. Switching was easy. The first 72 hours of quitting was a nightmare.

It's interesting how exactly this seems to parallel the rise of Hollywood in the 1920s-1960s sucking writers out of the rest of the world to work on scripts. The same risk/reward profile, the same looking down on the mass medium, the same blackhole of work if the rest of the massive project doesn't work out...

> Amazon.com: You have a new series known collectively as Edgeworks, where you are well into reprinting no less than 31 titles in 20 dual volumes. What makes them any different or special from the original editions, other than availability?

The given answer is mostly bullshit. The real answer seems to be that Ellison's undiagnosed & untreated bipolar disorder had given him both severe depression (so he couldn't write anything new & salable, even if he was not yet bedridden) and worsened his compulsive shopping habit, so his finances were in freefall. Ellison had made a lot of money over the years... and spent it all. Over the next decade even his _Babylon 5_ sinecure from his friend JMS would run out and he'd approach rock bottom (culminating in a firearm suicide attempt):

"Impulse spending sprees are as textbook bipolar as the manic episodes that lead to them, and Harlan’s growing inability to make good decisions began to decimate his savings. Despite a precipitous drop in income in the years after _Babylon 5_, Harlan kept making high-priced impulse purchases of comics, artwork, and collectibles. Susan would tell him, repeatedly, “We don’t have the money.” But the cash went out anyway, leading to panic when the bills came in, often paid with loans from friends, followed by the same reckless spending and crippled decision-making when the financial dust cleared. “He can’t control himself”, Susan said. “I’ve had to hide the credit cards.”"

See JMS's "Ellison Exegesis" in _Last Dangerous Visions_ last year (which points out that 1997 would see the last original volume from Ellison ever, so 1998 is at the top of the slippery slope where his life really began falling apart from bipolar).

I did some searching and I'm not sure the answer you remember exists. All I can find after a bunch of keyword + date-restricted searches in Google, Quora, & HN is Ibrahim Bashir (https://www.quora.com/Who-designed-UI-UX-and-developed-the-w...) who worked there 2003-2005 (https://www.quora.com/How-many-people-work-for-McMaster-Carr...) and who writes:

> McMaster-Carr has a Systems department which handles internal software development, including the website. I actually worked on a lot of the front-end functionality (among other things) during my tenure there and we had 1 person who was essentially the UX lead (I'm sure the team / function has expanded since then). > > The design philosophy for the website was heavily influence by Edward Tufte and myself and several folks from my engineering team were enrolled in his course to familiarize ourselves with key concepts. > > When I interviewed with McMaster-Carr I distinctly remember a Director who told me about building the first version of the website himself and not realizing it would become "a real thing" one day. I cannot for the life of me remember his name, but he was a sharp guy and I'm sure he's off doing great things. We walked into work one Monday morning, and his desk was empty and his whiteboard said "poof".

Given the timing and his description of what he did, it seems unlikely any other McMaster-Carr insider would have been writing about it, and this has to be what you are remembering, if anything.

(And https://iacollaborative.com/work/mcmaster-carr/ seems useless here.)

BTW, this is how it's going with the 'clever kids' and hiring for their cheating startup: https://x.com/im_roy_lee/status/1920206362996343261

    it's really fucking hard to find competent engineers. 

    i've interviewed about 20 cs majors who claim to be experts at typescript and only 3 of them have been able to explain what the purpose of useState is (not trolling).

(There are at least 2 ways this tweet is embarrassing.)

Unfortunately, the site was butchered a year or two ago with a redesign to make it look like a garden-variety Wordpress blog. A ton of content disappeared. The front page is reasonably dense, although one might also just call it cluttered, but the subpages are worse - sometimes a single paragraph lost in a sea of wrappers.

Thank you, ChatGPT, but these are all obvious questions to ask, and more valuable would be some answers.

He can have other motivations. Between 2020, 2024, Mackenzie Bezos & Laurene Powell Jobs, the deeply unimpressive philanthropy of the Buffett children, and his own divorce, a very rich philanthropist has excellent reasons to aim for the foundation being liquidated in his lifetime, and not handed off to administrators like, yes, the Ford Foundation or Harvard...

(And then, of course, given his enthusiasm for AI, there is a major question of whether 'keeping your powder dry' is a huge mistake - one way or the other.)